RPageStorageFile.cxx

Go to the documentation of this file.

/// \file RPageStorageFile.cxx
/// \ingroup NTuple
/// \author Jakob Blomer <jblomer@cern.ch>
/// \date 2019-11-25
 
/*************************************************************************
 * Copyright (C) 1995-2019, Rene Brun and Fons Rademakers.               *
 * All rights reserved.                                                  *
 *                                                                       *
 * For the licensing terms see $ROOTSYS/LICENSE.                         *
 * For the list of contributors see $ROOTSYS/README/CREDITS.             *
 *************************************************************************/
 
#include <ROOT/RCluster.hxx>
#include <ROOT/RClusterPool.hxx>
#include <ROOT/RLogger.hxx>
#include <ROOT/RNTupleDescriptor.hxx>
#include <ROOT/RNTupleModel.hxx>
#include <ROOT/RNTupleSerialize.hxx>
#include <ROOT/RNTupleZip.hxx>
#include <ROOT/RPage.hxx>
#include <ROOT/RPageAllocator.hxx>
#include <ROOT/RPagePool.hxx>
#include <ROOT/RPageStorageFile.hxx>
#include <ROOT/RRawFile.hxx>
#include <ROOT/RRawFileTFile.hxx>
#include <ROOT/RNTupleUtil.hxx>
 
#include <RVersion.h>
#include <TDirectory.h>
#include <TError.h>
#include <TVirtualStreamerInfo.h>
 
#include <algorithm>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <iterator>
#include <limits>
#include <utility>
 
#include <functional>
#include <mutex>
 
using ROOT::Experimental::Detail::RNTupleAtomicTimer;
using ROOT::Internal::MakeUninitArray;
using ROOT::Internal::RCluster;
using ROOT::Internal::RClusterPool;
using ROOT::Internal::RNTupleCompressor;
using ROOT::Internal::RNTupleDecompressor;
using ROOT::Internal::RNTupleFileWriter;
using ROOT::Internal::RNTupleSerializer;
using ROOT::Internal::ROnDiskPage;
using ROOT::Internal::ROnDiskPageMap;
using ROOT::Internal::RPagePool;
 
ROOT::Internal::RPageSinkFile::RPageSinkFile(std::string_view ntupleName, const ROOT::RNTupleWriteOptions &options)
   : RPagePersistentSink(ntupleName, options)
{
   EnableDefaultMetrics("RPageSinkFile");
   fFeatures.fCanMergePages = true;
}
 
ROOT::Internal::RPageSinkFile::RPageSinkFile(std::string_view ntupleName, std::string_view path,
                                             const ROOT::RNTupleWriteOptions &options)
   : RPageSinkFile(ntupleName, options)
{
   fWriter = RNTupleFileWriter::Recreate(ntupleName, path, RNTupleFileWriter::EContainerFormat::kTFile, options);
}
 
ROOT::Internal::RPageSinkFile::RPageSinkFile(std::string_view ntupleName, TDirectory &fileOrDirectory,
                                             const ROOT::RNTupleWriteOptions &options)
   : RPageSinkFile(ntupleName, options)
{
   fWriter = RNTupleFileWriter::Append(ntupleName, fileOrDirectory, options.GetMaxKeySize());
}
 
ROOT::Internal::RPageSinkFile::~RPageSinkFile() {}
 
void ROOT::Internal::RPageSinkFile::InitImpl(unsigned char *serializedHeader, std::uint32_t length)
{
   auto zipBuffer = MakeUninitArray<unsigned char>(length);
   auto szZipHeader =
      RNTupleCompressor::Zip(serializedHeader, length, GetWriteOptions().GetCompression(), zipBuffer.get());
   fWriter->WriteNTupleHeader(zipBuffer.get(), szZipHeader, length);
}
 
void ROOT::Internal::RPageSinkFile::UpdateSchema(const ROOT::Internal::RNTupleModelChangeset &changeset,
                                                 ROOT::NTupleSize_t firstEntry)
{
   RPagePersistentSink::UpdateSchema(changeset, firstEntry);
 
   auto fnAddStreamerInfo = [this](const ROOT::RFieldBase *field) {
      const TClass *cl = nullptr;
      if (auto classField = dynamic_cast<const RClassField *>(field)) {
         cl = classField->GetClass();
      } else if (auto streamerField = dynamic_cast<const RStreamerField *>(field)) {
         cl = streamerField->GetClass();
      }
      if (!cl)
         return;
 
      auto streamerInfo = cl->GetStreamerInfo(field->GetTypeVersion());
      if (!streamerInfo) {
         throw RException(R__FAIL(std::string("cannot get streamerInfo for ") + cl->GetName() + " [" +
                                  std::to_string(field->GetTypeVersion()) + "]"));
      }
      fInfosOfClassFields[streamerInfo->GetNumber()] = streamerInfo;
   };
 
   for (const auto field : changeset.fAddedFields) {
      fnAddStreamerInfo(field);
      for (const auto &subField : *field) {
         fnAddStreamerInfo(&subField);
      }
   }
}
 
inline ROOT::RNTupleLocator
ROOT::Internal::RPageSinkFile::WriteSealedPage(const RPageStorage::RSealedPage &sealedPage, std::size_t bytesPacked)
{
   std::uint64_t offsetData;
   {
      RNTupleAtomicTimer timer(fCounters->fTimeWallWrite, fCounters->fTimeCpuWrite);
      offsetData = fWriter->WriteBlob(sealedPage.GetBuffer(), sealedPage.GetBufferSize(), bytesPacked);
   }
 
   RNTupleLocator result;
   result.SetPosition(offsetData);
   result.SetNBytesOnStorage(sealedPage.GetDataSize());
   fCounters->fNPageCommitted.Inc();
   fCounters->fSzWritePayload.Add(sealedPage.GetBufferSize());
   fNBytesCurrentCluster += sealedPage.GetBufferSize();
   return result;
}
 
ROOT::RNTupleLocator
ROOT::Internal::RPageSinkFile::CommitPageImpl(ColumnHandle_t columnHandle, const ROOT::Internal::RPage &page)
{
   auto element = columnHandle.fColumn->GetElement();
   RPageStorage::RSealedPage sealedPage;
   {
      RNTupleAtomicTimer timer(fCounters->fTimeWallZip, fCounters->fTimeCpuZip);
      sealedPage = SealPage(page, *element);
   }
 
   fCounters->fSzZip.Add(page.GetNBytes());
   return WriteSealedPage(sealedPage, element->GetPackedSize(page.GetNElements()));
}
 
ROOT::RNTupleLocator ROOT::Internal::RPageSinkFile::CommitSealedPageImpl(ROOT::DescriptorId_t physicalColumnId,
                                                                         const RPageStorage::RSealedPage &sealedPage)
{
   const auto nBits = fDescriptorBuilder.GetDescriptor().GetColumnDescriptor(physicalColumnId).GetBitsOnStorage();
   const auto bytesPacked = (nBits * sealedPage.GetNElements() + 7) / 8;
   return WriteSealedPage(sealedPage, bytesPacked);
}
 
void ROOT::Internal::RPageSinkFile::CommitBatchOfPages(CommitBatch &batch, std::vector<RNTupleLocator> &locators)
{
   RNTupleAtomicTimer timer(fCounters->fTimeWallWrite, fCounters->fTimeCpuWrite);
 
   std::uint64_t offset = fWriter->ReserveBlob(batch.fSize, batch.fBytesPacked);
 
   locators.reserve(locators.size() + batch.fSealedPages.size());
 
   for (const auto *pagePtr : batch.fSealedPages) {
      fWriter->WriteIntoReservedBlob(pagePtr->GetBuffer(), pagePtr->GetBufferSize(), offset);
      RNTupleLocator locator;
      locator.SetPosition(offset);
      locator.SetNBytesOnStorage(pagePtr->GetDataSize());
      locators.push_back(locator);
      offset += pagePtr->GetBufferSize();
   }
 
   fCounters->fNPageCommitted.Add(batch.fSealedPages.size());
   fCounters->fSzWritePayload.Add(batch.fSize);
   fNBytesCurrentCluster += batch.fSize;
 
   batch.fSize = 0;
   batch.fBytesPacked = 0;
   batch.fSealedPages.clear();
}
 
std::vector<ROOT::RNTupleLocator>
ROOT::Internal::RPageSinkFile::CommitSealedPageVImpl(std::span<RPageStorage::RSealedPageGroup> ranges,
                                                     const std::vector<bool> &mask)
{
   const std::uint64_t maxKeySize = fOptions->GetMaxKeySize();
 
   CommitBatch batch{};
   std::vector<RNTupleLocator> locators;
 
   std::size_t iPage = 0;
   for (auto rangeIt = ranges.begin(); rangeIt != ranges.end(); ++rangeIt) {
      auto &range = *rangeIt;
      if (range.fFirst == range.fLast) {
         // Skip empty ranges, they might not have a physical column ID!
         continue;
      }
 
      const auto bitsOnStorage =
         fDescriptorBuilder.GetDescriptor().GetColumnDescriptor(range.fPhysicalColumnId).GetBitsOnStorage();
 
      for (auto sealedPageIt = range.fFirst; sealedPageIt != range.fLast; ++sealedPageIt, ++iPage) {
         if (!mask[iPage])
            continue;
 
         const auto bytesPacked = (bitsOnStorage * sealedPageIt->GetNElements() + 7) / 8;
 
         if (batch.fSize > 0 && batch.fSize + sealedPageIt->GetBufferSize() > maxKeySize) {
            /**
             * Adding this page would exceed maxKeySize. Since we always want to write into a single key
             * with vectorized writes, we commit the current set of pages before proceeding.
             * NOTE: we do this *before* checking if sealedPageIt->GetBufferSize() > maxKeySize to guarantee that
             * we always flush the current batch before doing an individual WriteBlob. This way we
             * preserve the assumption that a CommitBatch always contain a sequential set of pages.
             */
            CommitBatchOfPages(batch, locators);
         }
 
         if (sealedPageIt->GetBufferSize() > maxKeySize) {
            // This page alone is bigger than maxKeySize: save it by itself, since it will need to be
            // split into multiple keys.
 
            // Since this check implies the previous check on batchSize + newSize > maxSize, we should
            // already have committed the current batch before writing this page.
            assert(batch.fSize == 0);
 
            std::uint64_t offset =
               fWriter->WriteBlob(sealedPageIt->GetBuffer(), sealedPageIt->GetBufferSize(), bytesPacked);
            RNTupleLocator locator;
            locator.SetPosition(offset);
            locator.SetNBytesOnStorage(sealedPageIt->GetDataSize());
            locators.push_back(locator);
 
            fCounters->fNPageCommitted.Inc();
            fCounters->fSzWritePayload.Add(sealedPageIt->GetBufferSize());
            fNBytesCurrentCluster += sealedPageIt->GetBufferSize();
 
         } else {
            batch.fSealedPages.emplace_back(&(*sealedPageIt));
            batch.fSize += sealedPageIt->GetBufferSize();
            batch.fBytesPacked += bytesPacked;
         }
      }
   }
 
   if (batch.fSize > 0) {
      CommitBatchOfPages(batch, locators);
   }
 
   return locators;
}
 
std::uint64_t ROOT::Internal::RPageSinkFile::StageClusterImpl()
{
   auto result = fNBytesCurrentCluster;
   fNBytesCurrentCluster = 0;
   return result;
}
 
ROOT::RNTupleLocator
ROOT::Internal::RPageSinkFile::CommitClusterGroupImpl(unsigned char *serializedPageList, std::uint32_t length)
{
   auto bufPageListZip = MakeUninitArray<unsigned char>(length);
   auto szPageListZip =
      RNTupleCompressor::Zip(serializedPageList, length, GetWriteOptions().GetCompression(), bufPageListZip.get());
 
   RNTupleLocator result;
   result.SetNBytesOnStorage(szPageListZip);
   result.SetPosition(fWriter->WriteBlob(bufPageListZip.get(), szPageListZip, length));
   return result;
}
 
void ROOT::Internal::RPageSinkFile::CommitDatasetImpl(unsigned char *serializedFooter, std::uint32_t length)
{
   // Add the streamer info records from streamer fields: because of runtime polymorphism we may need to add additional
   // types not covered by the type names of the class fields
   for (const auto &extraTypeInfo : fDescriptorBuilder.GetDescriptor().GetExtraTypeInfoIterable()) {
      if (extraTypeInfo.GetContentId() != EExtraTypeInfoIds::kStreamerInfo)
         continue;
      // Ideally, we would avoid deserializing the streamer info records of the streamer fields that we just serialized.
      // However, this happens only once at the end of writing and only when streamer fields are used, so the
      // preference here is for code simplicity.
      fInfosOfClassFields.merge(RNTupleSerializer::DeserializeStreamerInfos(extraTypeInfo.GetContent()).Unwrap());
   }
   fWriter->UpdateStreamerInfos(fInfosOfClassFields);
 
   auto bufFooterZip = MakeUninitArray<unsigned char>(length);
   auto szFooterZip =
      RNTupleCompressor::Zip(serializedFooter, length, GetWriteOptions().GetCompression(), bufFooterZip.get());
   fWriter->WriteNTupleFooter(bufFooterZip.get(), szFooterZip, length);
   fWriter->Commit(GetWriteOptions().GetCompression());
}
 
////////////////////////////////////////////////////////////////////////////////
 
ROOT::Internal::RPageSourceFile::RPageSourceFile(std::string_view ntupleName, const ROOT::RNTupleReadOptions &opts)
   : RPageSource(ntupleName, opts),
     fClusterPool(
        std::make_unique<RClusterPool>(*this, ROOT::Internal::RNTupleReadOptionsManip::GetClusterBunchSize(opts)))
{
   EnableDefaultMetrics("RPageSourceFile");
}
 
ROOT::Internal::RPageSourceFile::RPageSourceFile(std::string_view ntupleName,
                                                 std::unique_ptr<ROOT::Internal::RRawFile> file,
                                                 const ROOT::RNTupleReadOptions &options)
   : RPageSourceFile(ntupleName, options)
{
   fFile = std::move(file);
   R__ASSERT(fFile);
   fReader = ROOT::Internal::RMiniFileReader(fFile.get());
}
 
ROOT::Internal::RPageSourceFile::RPageSourceFile(std::string_view ntupleName, std::string_view path,
                                                 const ROOT::RNTupleReadOptions &options)
   : RPageSourceFile(ntupleName, ROOT::Internal::RRawFile::Create(path), options)
{
}
 
std::unique_ptr<ROOT::Internal::RPageSourceFile>
ROOT::Internal::RPageSourceFile::CreateFromAnchor(const RNTuple &anchor, const ROOT::RNTupleReadOptions &options)
{
   if (!anchor.fFile)
      throw RException(R__FAIL("This RNTuple object was not streamed from a ROOT file (TFile or descendant)"));
 
   std::unique_ptr<ROOT::Internal::RRawFile> rawFile;
   // For local TFiles, TDavixFile, and TNetXNGFile, we want to open a new RRawFile to take advantage of the faster
   // reading. We check the exact class name to avoid classes inheriting in ROOT (for example TMemFile) or in
   // experiment frameworks.
   std::string className = anchor.fFile->IsA()->GetName();
   auto url = anchor.fFile->GetEndpointUrl();
   auto protocol = std::string(url->GetProtocol());
   if (className == "TFile") {
      rawFile = ROOT::Internal::RRawFile::Create(url->GetFile());
   } else if (className == "TDavixFile" || className == "TNetXNGFile") {
      rawFile = ROOT::Internal::RRawFile::Create(url->GetUrl());
   } else {
      rawFile.reset(new ROOT::Internal::RRawFileTFile(anchor.fFile));
   }
 
   auto pageSource = std::make_unique<RPageSourceFile>("", std::move(rawFile), options);
   pageSource->fAnchor = anchor;
   pageSource->fNTupleName = pageSource->fDescriptorBuilder.GetDescriptor().GetName();
   return pageSource;
}
 
ROOT::Internal::RPageSourceFile::~RPageSourceFile() = default;
 
void ROOT::Internal::RPageSourceFile::LoadStructureImpl()
{
   // If we constructed the page source with (ntuple name, path), we need to find the anchor first.
   // Otherwise, the page source was created by OpenFromAnchor()
   if (!fAnchor) {
      fAnchor = fReader.GetNTuple(fNTupleName).Unwrap();
   }
   fReader.SetMaxKeySize(fAnchor->GetMaxKeySize());
 
   fDescriptorBuilder.SetVersion(fAnchor->GetVersionEpoch(), fAnchor->GetVersionMajor(), fAnchor->GetVersionMinor(),
                                 fAnchor->GetVersionPatch());
   fDescriptorBuilder.SetOnDiskHeaderSize(fAnchor->GetNBytesHeader());
   fDescriptorBuilder.AddToOnDiskFooterSize(fAnchor->GetNBytesFooter());
 
   // Reserve enough space for the compressed and the uncompressed header/footer (see AttachImpl)
   const auto bufSize = fAnchor->GetNBytesHeader() + fAnchor->GetNBytesFooter() +
                        std::max(fAnchor->GetLenHeader(), fAnchor->GetLenFooter());
   fStructureBuffer.fBuffer = MakeUninitArray<unsigned char>(bufSize);
   fStructureBuffer.fPtrHeader = fStructureBuffer.fBuffer.get();
   fStructureBuffer.fPtrFooter = fStructureBuffer.fBuffer.get() + fAnchor->GetNBytesHeader();
 
   auto readvLimits = fFile->GetReadVLimits();
   // Never try to vectorize reads to a split key
   readvLimits.fMaxSingleSize = std::min<size_t>(readvLimits.fMaxSingleSize, fAnchor->GetMaxKeySize());
 
   if ((readvLimits.fMaxReqs < 2) ||
       (std::max(fAnchor->GetNBytesHeader(), fAnchor->GetNBytesFooter()) > readvLimits.fMaxSingleSize) ||
       (fAnchor->GetNBytesHeader() + fAnchor->GetNBytesFooter() > readvLimits.fMaxTotalSize)) {
      RNTupleAtomicTimer timer(fCounters->fTimeWallRead, fCounters->fTimeCpuRead);
      fReader.ReadBuffer(fStructureBuffer.fPtrHeader, fAnchor->GetNBytesHeader(), fAnchor->GetSeekHeader());
      fReader.ReadBuffer(fStructureBuffer.fPtrFooter, fAnchor->GetNBytesFooter(), fAnchor->GetSeekFooter());
      fCounters->fNRead.Add(2);
   } else {
      RNTupleAtomicTimer timer(fCounters->fTimeWallRead, fCounters->fTimeCpuRead);
      R__ASSERT(fAnchor->GetNBytesHeader() < std::numeric_limits<std::size_t>::max());
      R__ASSERT(fAnchor->GetNBytesFooter() < std::numeric_limits<std::size_t>::max());
      ROOT::Internal::RRawFile::RIOVec readRequests[2] = {{fStructureBuffer.fPtrHeader, fAnchor->GetSeekHeader(),
                                                           static_cast<std::size_t>(fAnchor->GetNBytesHeader()), 0},
                                                          {fStructureBuffer.fPtrFooter, fAnchor->GetSeekFooter(),
                                                           static_cast<std::size_t>(fAnchor->GetNBytesFooter()), 0}};
      fFile->ReadV(readRequests, 2);
      fCounters->fNReadV.Inc();
   }
}
 
ROOT::RNTupleDescriptor ROOT::Internal::RPageSourceFile::AttachImpl(RNTupleSerializer::EDescriptorDeserializeMode mode)
{
   auto unzipBuf = reinterpret_cast<unsigned char *>(fStructureBuffer.fPtrFooter) + fAnchor->GetNBytesFooter();
 
   RNTupleDecompressor::Unzip(fStructureBuffer.fPtrHeader, fAnchor->GetNBytesHeader(), fAnchor->GetLenHeader(),
                              unzipBuf);
   RNTupleSerializer::DeserializeHeader(unzipBuf, fAnchor->GetLenHeader(), fDescriptorBuilder);
 
   RNTupleDecompressor::Unzip(fStructureBuffer.fPtrFooter, fAnchor->GetNBytesFooter(), fAnchor->GetLenFooter(),
                              unzipBuf);
   RNTupleSerializer::DeserializeFooter(unzipBuf, fAnchor->GetLenFooter(), fDescriptorBuilder);
 
   auto desc = fDescriptorBuilder.MoveDescriptor();
 
   std::vector<unsigned char> buffer;
   for (const auto &cgDesc : desc.GetClusterGroupIterable()) {
      buffer.resize(std::max<size_t>(buffer.size(),
                                     cgDesc.GetPageListLength() + cgDesc.GetPageListLocator().GetNBytesOnStorage()));
      auto *zipBuffer = buffer.data() + cgDesc.GetPageListLength();
      fReader.ReadBuffer(zipBuffer, cgDesc.GetPageListLocator().GetNBytesOnStorage(),
                         cgDesc.GetPageListLocator().GetPosition<std::uint64_t>());
      RNTupleDecompressor::Unzip(zipBuffer, cgDesc.GetPageListLocator().GetNBytesOnStorage(),
                                 cgDesc.GetPageListLength(), buffer.data());
 
      RNTupleSerializer::DeserializePageList(buffer.data(), cgDesc.GetPageListLength(), cgDesc.GetId(), desc, mode);
   }
 
   // For the page reads, we rely on the I/O scheduler to define the read requests
   fFile->SetBuffering(false);
 
   return desc;
}
 
void ROOT::Internal::RPageSourceFile::LoadSealedPage(ROOT::DescriptorId_t physicalColumnId,
                                                     RNTupleLocalIndex localIndex, RSealedPage &sealedPage)
{
   const auto clusterId = localIndex.GetClusterId();
 
   ROOT::RClusterDescriptor::RPageInfo pageInfo;
   {
      auto descriptorGuard = GetSharedDescriptorGuard();
      const auto &clusterDescriptor = descriptorGuard->GetClusterDescriptor(clusterId);
      pageInfo = clusterDescriptor.GetPageRange(physicalColumnId).Find(localIndex.GetIndexInCluster());
   }
 
   sealedPage.SetBufferSize(pageInfo.GetLocator().GetNBytesOnStorage() + pageInfo.HasChecksum() * kNBytesPageChecksum);
   sealedPage.SetNElements(pageInfo.GetNElements());
   sealedPage.SetHasChecksum(pageInfo.HasChecksum());
   if (!sealedPage.GetBuffer())
      return;
   if (pageInfo.GetLocator().GetType() != RNTupleLocator::kTypePageZero) {
      fReader.ReadBuffer(const_cast<void *>(sealedPage.GetBuffer()), sealedPage.GetBufferSize(),
                         pageInfo.GetLocator().GetPosition<std::uint64_t>());
   } else {
      assert(!pageInfo.HasChecksum());
      memcpy(const_cast<void *>(sealedPage.GetBuffer()), ROOT::Internal::RPage::GetPageZeroBuffer(),
             sealedPage.GetBufferSize());
   }
 
   sealedPage.VerifyChecksumIfEnabled().ThrowOnError();
}
 
ROOT::Internal::RPageRef ROOT::Internal::RPageSourceFile::LoadPageImpl(ColumnHandle_t columnHandle,
                                                                       const RClusterInfo &clusterInfo,
                                                                       ROOT::NTupleSize_t idxInCluster)
{
   const auto columnId = columnHandle.fPhysicalId;
   const auto clusterId = clusterInfo.fClusterId;
   const auto pageInfo = clusterInfo.fPageInfo;
 
   const auto element = columnHandle.fColumn->GetElement();
   const auto elementSize = element->GetSize();
   const auto elementInMemoryType = element->GetIdentifier().fInMemoryType;
 
   if (pageInfo.GetLocator().GetType() == RNTupleLocator::kTypePageZero) {
      auto pageZero = fPageAllocator->NewPage(elementSize, pageInfo.GetNElements());
      pageZero.GrowUnchecked(pageInfo.GetNElements());
      memset(pageZero.GetBuffer(), 0, pageZero.GetNBytes());
      pageZero.SetWindow(clusterInfo.fColumnOffset + pageInfo.GetFirstElementIndex(),
                         ROOT::Internal::RPage::RClusterInfo(clusterId, clusterInfo.fColumnOffset));
      return fPagePool.RegisterPage(std::move(pageZero), RPagePool::RKey{columnId, elementInMemoryType});
   }
 
   RSealedPage sealedPage;
   sealedPage.SetNElements(pageInfo.GetNElements());
   sealedPage.SetHasChecksum(pageInfo.HasChecksum());
   sealedPage.SetBufferSize(pageInfo.GetLocator().GetNBytesOnStorage() + pageInfo.HasChecksum() * kNBytesPageChecksum);
   std::unique_ptr<unsigned char[]> directReadBuffer; // only used if cluster pool is turned off
 
   if (fOptions.GetClusterCache() == ROOT::RNTupleReadOptions::EClusterCache::kOff) {
      directReadBuffer = MakeUninitArray<unsigned char>(sealedPage.GetBufferSize());
      {
         RNTupleAtomicTimer timer(fCounters->fTimeWallRead, fCounters->fTimeCpuRead);
         fReader.ReadBuffer(directReadBuffer.get(), sealedPage.GetBufferSize(),
                            pageInfo.GetLocator().GetPosition<std::uint64_t>());
      }
      fCounters->fNPageRead.Inc();
      fCounters->fNRead.Inc();
      fCounters->fSzReadPayload.Add(sealedPage.GetBufferSize());
      sealedPage.SetBuffer(directReadBuffer.get());
   } else {
      if (!fCurrentCluster || (fCurrentCluster->GetId() != clusterId) || !fCurrentCluster->ContainsColumn(columnId))
         fCurrentCluster = fClusterPool->GetCluster(clusterId, fActivePhysicalColumns.ToColumnSet());
      R__ASSERT(fCurrentCluster->ContainsColumn(columnId));
 
      auto cachedPageRef =
         fPagePool.GetPage(RPagePool::RKey{columnId, elementInMemoryType}, RNTupleLocalIndex(clusterId, idxInCluster));
      if (!cachedPageRef.Get().IsNull())
         return cachedPageRef;
 
      ROnDiskPage::Key key(columnId, pageInfo.GetPageNumber());
      auto onDiskPage = fCurrentCluster->GetOnDiskPage(key);
      R__ASSERT(onDiskPage && (sealedPage.GetBufferSize() == onDiskPage->GetSize()));
      sealedPage.SetBuffer(onDiskPage->GetAddress());
   }
 
   ROOT::Internal::RPage newPage;
   {
      RNTupleAtomicTimer timer(fCounters->fTimeWallUnzip, fCounters->fTimeCpuUnzip);
      newPage = UnsealPage(sealedPage, *element).Unwrap();
      fCounters->fSzUnzip.Add(elementSize * pageInfo.GetNElements());
   }
 
   newPage.SetWindow(clusterInfo.fColumnOffset + pageInfo.GetFirstElementIndex(),
                     ROOT::Internal::RPage::RClusterInfo(clusterId, clusterInfo.fColumnOffset));
   fCounters->fNPageUnsealed.Inc();
   return fPagePool.RegisterPage(std::move(newPage), RPagePool::RKey{columnId, elementInMemoryType});
}
 
std::unique_ptr<ROOT::Internal::RPageSource> ROOT::Internal::RPageSourceFile::CloneImpl() const
{
   auto clone = new RPageSourceFile(fNTupleName, fOptions);
   clone->fFile = fFile->Clone();
   clone->fReader = ROOT::Internal::RMiniFileReader(clone->fFile.get());
   return std::unique_ptr<RPageSourceFile>(clone);
}
 
std::unique_ptr<ROOT::Internal::RCluster>
ROOT::Internal::RPageSourceFile::PrepareSingleCluster(const RCluster::RKey &clusterKey,
                                                      std::vector<ROOT::Internal::RRawFile::RIOVec> &readRequests)
{
   struct ROnDiskPageLocator {
      ROOT::DescriptorId_t fColumnId = 0;
      ROOT::NTupleSize_t fPageNo = 0;
      std::uint64_t fOffset = 0;
      std::uint64_t fSize = 0;
      std::size_t fBufPos = 0;
   };
 
   std::vector<ROnDiskPageLocator> onDiskPages;
   auto activeSize = 0;
   auto pageZeroMap = std::make_unique<ROnDiskPageMap>();
   PrepareLoadCluster(clusterKey, *pageZeroMap,
                      [&](ROOT::DescriptorId_t physicalColumnId, ROOT::NTupleSize_t pageNo,
                          const ROOT::RClusterDescriptor::RPageInfo &pageInfo) {
                         const auto &pageLocator = pageInfo.GetLocator();
                         if (pageLocator.GetType() == RNTupleLocator::kTypeUnknown)
                            throw RException(R__FAIL("tried to read a page with an unknown locator"));
                         const auto nBytes =
                            pageLocator.GetNBytesOnStorage() + pageInfo.HasChecksum() * kNBytesPageChecksum;
                         activeSize += nBytes;
                         onDiskPages.push_back(
                            {physicalColumnId, pageNo, pageLocator.GetPosition<std::uint64_t>(), nBytes, 0});
                      });
 
   // Linearize the page requests by file offset
   std::sort(onDiskPages.begin(), onDiskPages.end(),
             [](const ROnDiskPageLocator &a, const ROnDiskPageLocator &b) { return a.fOffset < b.fOffset; });
 
   // In order to coalesce close-by pages, we collect the sizes of the gaps between pages on disk.  We then order
   // the gaps by size, sum them up and find a cutoff for the largest gap that we tolerate when coalescing pages.
   // The size of the cutoff is given by the fraction of extra bytes we are willing to read in order to reduce
   // the number of read requests.  We thus schedule the lowest number of requests given a tolerable fraction
   // of extra bytes.
   // TODO(jblomer): Eventually we may want to select the parameter at runtime according to link latency and speed,
   // memory consumption, device block size.
   float maxOverhead = 0.25 * float(activeSize);
   std::vector<std::size_t> gaps;
   if (onDiskPages.size())
      gaps.reserve(onDiskPages.size() - 1);
   for (unsigned i = 1; i < onDiskPages.size(); ++i) {
      std::int64_t gap =
         static_cast<int64_t>(onDiskPages[i].fOffset) - (onDiskPages[i - 1].fSize + onDiskPages[i - 1].fOffset);
      gaps.emplace_back(std::max(gap, std::int64_t(0)));
      // If the pages overlap, substract the overlapped bytes from `activeSize`
      activeSize += std::min(gap, std::int64_t(0));
   }
   std::sort(gaps.begin(), gaps.end());
   std::size_t gapCut = 0;
   std::size_t currentGap = 0;
   float szExtra = 0.0;
   for (auto g : gaps) {
      if (g != currentGap) {
         gapCut = currentGap;
         currentGap = g;
      }
      szExtra += g;
      if (szExtra > maxOverhead)
         break;
   }
 
   // In a first step, we coalesce the read requests and calculate the cluster buffer size.
   // In a second step, we'll fix-up the memory destinations for the read calls given the
   // address of the allocated buffer.  We must not touch, however, the read requests from previous
   // calls to PrepareSingleCluster()
   const auto currentReadRequestIdx = readRequests.size();
 
   ROOT::Internal::RRawFile::RIOVec req;
   // To simplify the first loop iteration, pretend an empty request starting at the first page's fOffset.
   if (!onDiskPages.empty())
      req.fOffset = onDiskPages[0].fOffset;
   std::size_t szPayload = 0;
   std::size_t szOverhead = 0;
   const std::uint64_t maxKeySize = fReader.GetMaxKeySize();
   for (auto &s : onDiskPages) {
      R__ASSERT(s.fSize > 0);
      const std::int64_t readUpTo = req.fOffset + req.fSize;
      // Note: byte ranges of pages may overlap
      const std::uint64_t overhead = std::max(static_cast<std::int64_t>(s.fOffset) - readUpTo, std::int64_t(0));
      const std::uint64_t extent = std::max(static_cast<std::int64_t>(s.fOffset + s.fSize) - readUpTo, std::int64_t(0));
      if (req.fSize + extent < maxKeySize && overhead <= gapCut) {
         szPayload += (extent - overhead);
         szOverhead += overhead;
         s.fBufPos = reinterpret_cast<intptr_t>(req.fBuffer) + s.fOffset - req.fOffset;
         req.fSize += extent;
         continue;
      }
 
      // close the current request and open new one
      if (req.fSize > 0)
         readRequests.emplace_back(req);
 
      req.fBuffer = reinterpret_cast<unsigned char *>(req.fBuffer) + req.fSize;
      s.fBufPos = reinterpret_cast<intptr_t>(req.fBuffer);
 
      szPayload += s.fSize;
      req.fOffset = s.fOffset;
      req.fSize = s.fSize;
   }
   readRequests.emplace_back(req);
   fCounters->fSzReadPayload.Add(szPayload);
   fCounters->fSzReadOverhead.Add(szOverhead);
 
   // Register the on disk pages in a page map
   auto buffer = new unsigned char[reinterpret_cast<intptr_t>(req.fBuffer) + req.fSize];
   auto pageMap = std::make_unique<ROOT::Internal::ROnDiskPageMapHeap>(std::unique_ptr<unsigned char[]>(buffer));
   for (const auto &s : onDiskPages) {
      ROnDiskPage::Key key(s.fColumnId, s.fPageNo);
      pageMap->Register(key, ROnDiskPage(buffer + s.fBufPos, s.fSize));
   }
   fCounters->fNPageRead.Add(onDiskPages.size());
   for (auto i = currentReadRequestIdx; i < readRequests.size(); ++i) {
      readRequests[i].fBuffer = buffer + reinterpret_cast<intptr_t>(readRequests[i].fBuffer);
   }
 
   auto cluster = std::make_unique<RCluster>(clusterKey.fClusterId);
   cluster->Adopt(std::move(pageMap));
   cluster->Adopt(std::move(pageZeroMap));
   for (auto colId : clusterKey.fPhysicalColumnSet)
      cluster->SetColumnAvailable(colId);
   return cluster;
}
 
std::vector<std::unique_ptr<ROOT::Internal::RCluster>>
ROOT::Internal::RPageSourceFile::LoadClusters(std::span<RCluster::RKey> clusterKeys)
{
   fCounters->fNClusterLoaded.Add(clusterKeys.size());
 
   std::vector<std::unique_ptr<ROOT::Internal::RCluster>> clusters;
   std::vector<ROOT::Internal::RRawFile::RIOVec> readRequests;
 
   clusters.reserve(clusterKeys.size());
   for (auto key : clusterKeys) {
      clusters.emplace_back(PrepareSingleCluster(key, readRequests));
   }
 
   auto nReqs = readRequests.size();
   auto readvLimits = fFile->GetReadVLimits();
   // We never want to do vectorized reads of split blobs, so we limit our single size to maxKeySize.
   readvLimits.fMaxSingleSize = std::min<size_t>(readvLimits.fMaxSingleSize, fReader.GetMaxKeySize());
 
   int iReq = 0;
   while (nReqs > 0) {
      auto nBatch = std::min(nReqs, readvLimits.fMaxReqs);
 
      if (readvLimits.HasSizeLimit()) {
         std::uint64_t totalSize = 0;
         for (std::size_t i = 0; i < nBatch; ++i) {
            if (readRequests[iReq + i].fSize > readvLimits.fMaxSingleSize) {
               nBatch = i;
               break;
            }
 
            totalSize += readRequests[iReq + i].fSize;
            if (totalSize > readvLimits.fMaxTotalSize) {
               nBatch = i;
               break;
            }
         }
      }
 
      if (nBatch <= 1) {
         nBatch = 1;
         RNTupleAtomicTimer timer(fCounters->fTimeWallRead, fCounters->fTimeCpuRead);
         fReader.ReadBuffer(readRequests[iReq].fBuffer, readRequests[iReq].fSize, readRequests[iReq].fOffset);
      } else {
         RNTupleAtomicTimer timer(fCounters->fTimeWallRead, fCounters->fTimeCpuRead);
         fFile->ReadV(&readRequests[iReq], nBatch);
      }
      fCounters->fNReadV.Inc();
      fCounters->fNRead.Add(nBatch);
 
      iReq += nBatch;
      nReqs -= nBatch;
   }
 
   return clusters;
}